measure provider prompt cache hits

cmd/olliesrv/internal/agent/agent.go

@@ -21,18 +21,18 @@ import (
 // a new Agent is built from the new agent config while the session
 // host remains stable.
 type Agent struct {
-	history        *History
-	runtime        *Runtime
-	name           string // display name (defaults to id)
-	profile        string      // config profile name (e.g. "default")
-	agentsDir      string
-	systemPrompt   string   // system prompt for /agent reloads
-	envBlock       string   // environment block for /agent reloads
-	newToolServer  func() *toolsrv.Conn
-	newBackend     func(string) (backend.Backend, error)
-	currentAction  atomic.Pointer[actionHandle]
-	warnedContext  bool
-	resultCache    sync.Map
+	history       *History
+	runtime       *Runtime
+	name          string // display name (defaults to id)
+	profile       string // config profile name (e.g. "default")
+	agentsDir     string
+	systemPrompt  string // system prompt for /agent reloads
+	envBlock      string // environment block for /agent reloads
+	newToolServer func() *toolsrv.Conn
+	newBackend    func(string) (backend.Backend, error)
+	currentAction atomic.Pointer[actionHandle]
+	warnedContext bool
+	resultCache   sync.Map
 
 	// Execution state — owned by the agent, protected by stateMu.
 	state         string // "idle", "thinking", "calling: <tool>"
@@ -43,10 +43,10 @@ type Agent struct {
 	Feed          Feed   // streaming input gate (change-detecting)
 	toolCallCount atomic.Int64
 	pendingInject atomic.Pointer[string]
-	submitMu     sync.Mutex // serializes Submit calls (commands + turns)
-	stateMu      sync.RWMutex
-	signalMu     sync.Mutex
-	signalCh     chan struct{} // closed on state change; replaced with a fresh channel
+	submitMu      sync.Mutex // serializes Submit calls (commands + turns)
+	stateMu       sync.RWMutex
+	signalMu      sync.Mutex
+	signalCh      chan struct{} // closed on state change; replaced with a fresh channel
 
 	// Event handler — receives events from this agent.
 	output          EventHandler
@@ -96,6 +96,7 @@ type UsageStats struct {
 	Estimated                bool
 	LastTurnCostUSD          float64
 	SessionCostUSD           float64
+	CacheHitRatio            float64
 }
 
 // Usage returns the agent's usage statistics.
@@ -112,6 +113,7 @@ func (ag *Agent) Usage() *UsageStats {
 		Estimated:                ag.history.Estimated,
 		LastTurnCostUSD:          ag.history.LastTurnCostUSD,
 		SessionCostUSD:           ag.history.SessionCostUSD,
+		CacheHitRatio:            ag.history.cacheHitRatio(),
 	}
 }
 
@@ -491,7 +493,6 @@ const (
 	WatchFeed  = "feed"
 )
 
-
 // SetSessionEnv injects session env vars into the execute server.
 func (ag *Agent) SetSessionEnv(sessionID string) {
 	if ag.runtime == nil || ag.runtime.ToolServer == nil {
@@ -670,26 +671,24 @@ func ConsumeFeed(ctx context.Context, ag *Agent) {
 	}
 }
 
-
-
 // AgentParams holds the runtime dependencies for constructing a new Agent.
 // Not to be confused with AgentConfig, which is the on-disk JSON schema.
 type AgentParams struct {
-	ID             string // unique agent identity (uname)
-	SessionID      string // id of session the agent belongs to
-	Profile        string // config profile name (e.g. "default" → agents/default.json)
-	History        *History
-	Runtime        *Runtime
-	AgentsDir      string
-	Cwd            string // working directory for tool execution
-	SystemPrompt   string
-	EnvBlock       string
-	NewToolServer  func() *toolsrv.Conn
-	NewBackend     func(string) (backend.Backend, error)
-	Log            *olog.Logger
-	StartupMsgs    []string
-	Save           func()
-	Flush          func()
+	ID            string // unique agent identity (uname)
+	SessionID     string // id of session the agent belongs to
+	Profile       string // config profile name (e.g. "default" → agents/default.json)
+	History       *History
+	Runtime       *Runtime
+	AgentsDir     string
+	Cwd           string // working directory for tool execution
+	SystemPrompt  string
+	EnvBlock      string
+	NewToolServer func() *toolsrv.Conn
+	NewBackend    func(string) (backend.Backend, error)
+	Log           *olog.Logger
+	StartupMsgs   []string
+	Save          func()
+	Flush         func()
 }
 
 // NewAgent constructs an Agent from the given configuration.

cmd/olliesrv/internal/agent/history.go

@@ -151,6 +151,13 @@ func (s *History) addUsage(u backend.Usage, estimated bool) {
 	}
 }
 
+func (s *History) cacheHitRatio() float64 {
+	total := s.TotalInputTokens + s.TotalCachedInputTokens
+	if total == 0 {
+		return 0
+	}
+	return float64(s.TotalCachedInputTokens) / float64(total)
+}
 func (s *History) resetTurnAccumulators() {
 	s.turnInputTokens = 0
 	s.turnCachedTokens = 0

cmd/olliesrv/internal/agent/turn.go

@@ -165,13 +165,15 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 			var in, out, est, cached, creation int
 			var costUSD float64
 			fmt.Sscanf(ev.Content, "%d %d %d %g %d %d", &in, &out, &est, &costUSD, &cached, &creation)
-			ag.history.addUsage(backend.Usage{
+			usage := backend.Usage{
 				InputTokens:         in,
 				CachedInputTokens:   cached,
 				CacheCreationTokens: creation,
 				OutputTokens:        out,
 				CostUSD:             costUSD,
-			}, est != 0)
+			}
+			ag.history.addUsage(usage, est != 0)
+			ag.log.Debug("usage: model=%s input=%d cached=%d cache_create=%d output=%d hit_ratio=%.3f cost=$%.6f", ag.runtime.Backend.Model(), in, cached, creation, out, ag.history.cacheHitRatio(), costUSD)
 			ag.notifyChange()
 		}
 		if origOutput != nil {